Wrapper methods are hyper-parameter selection methods that

Wrapper methods are hyper-parameter selection methods that involve training multiple models with different subsets of features and selecting the best subset based on performance metrics. They are particularly useful when the underlying learning algorithms are "black boxes," meaning their internal workings are not easily interpretable or understood.
Option A is incorrect because wrapper methods can be computationally intensive since they involve training multiple models.
Option B is incorrect because whether wrapper methods are prone to overfitting depends on their implementation and the data.
Option D is incorrect because wrapper methods can be valuable in certain scenarios, especially when interpretability is not a primary concern.
Therefore, the correct answer is Option C: are useful mainly when the learning machines are "black boxes".